Generative Semantic Communication (GSC)

• Generative Semantic Communication (GSC) is a paradigm that transmits compact, task-relevant semantic information for content regeneration using advanced generative models.
• It decouples semantic encoding and decoding, enabling the use of diffusion models, LLMs, and knowledge graphs to optimize performance under various channel constraints.
• Empirical studies show that GSC can improve metrics such as PSNR, energy efficiency, and retrieval accuracy across modalities like images, text, video, and metaverse applications.

Generative Semantic Communication (GSC) is a semantic-communication paradigm in which the transmitter conveys compact semantic information and the receiver uses generative models to regenerate task-aligned content, rather than attempting bit-perfect or pixel-perfect recovery. In the current literature, this shift is explicitly framed as moving from “information recovery” to “information regeneration,” with the receiver directly generating the desired content from coded semantic information, shared knowledge, or task-conditioned embeddings (Ren et al., 2024). In AGI-driven formulations, GSC is further defined for “general-sense” tasks and human-friendly outputs, with separate task-relevant and perceptual semantic subgraphs embedded and transmitted under a rate–distortion–perception trade-off (Yuan et al., 21 Apr 2025).

1. Conceptual foundations

The point of departure for GSC is the limitation of Shannon-style communication as a universal design template for intelligent services. In that template, source and channel coding target faithful symbol recovery regardless of meaning. Semantic communication relaxes this requirement by transmitting only information relevant to a task; GSC extends it by letting a generative model reconstruct or recreate the final output from compact semantics, prompts, latent variables, or structured knowledge (Ren et al., 2024).

Several papers formalize this transition differently, but the common structure is stable. In the LLM-native formulation, semantic encoding and regeneration are written as

$z = S(x)\in\{1,\dots,M\}^L, \qquad \hat y = R(z),$

where the transmitted object is a sequence of token indices and the receiver regenerates task-specific content (Ren et al., 2024). In AGI-driven formulations, the transmitter extracts both task and perceptual embeddings, and semantic distortion is measured in embedding space, for example

$$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$

while perceptual quality is constrained separately (Yuan et al., 21 Apr 2025).

A second foundational theme is that GSC is not restricted to a single modality or a single semantic object. Across the literature, the transmitted carrier may be a semantic map, a keypoint set, a prompt text, a diffusion latent, a task instruction, a set of knowledge-graph node IDs, or a low-dimensional embedding sequence (Yang et al., 2023, Hu et al., 24 Jun 2025, Fan et al., 5 Sep 2025). This suggests that GSC is better understood as a regeneration-centric design principle than as one standardized encoder–decoder stack.

A third theme is decoupling. Multiple works stress that the semantic encoder and semantic decoder can be optimized independently, in contrast to monolithic end-to-end semantic autoencoders. This decoupling is presented as a route to explainability, modularity, and robustness, especially when the decoder can use strong priors from diffusion models, LLMs, or knowledge bases to compensate for incomplete or corrupted semantics (Yang et al., 2023, Ren et al., 2023).

2. Generative models and canonical system architectures

The architectural core of GSC spans several generative model families. A broad survey identifies three “classical GAI” families—variational autoencoders, generative adversarial networks, and diffusion models—as foundational SemCom enablers, and then introduces an LLM-based system with an “Understanding AI Agent” at the transmitter and a “Generating AI Agent” at the receiver (Ren et al., 2024). In that system, the transmitter perceives raw multimodal data, channel state, and task requirement, and produces a discrete understanding embedding; the receiver directly regenerates images, video clips, or text summaries from the received embedding (Ren et al., 2024).

Image-centric GSC systems frequently combine a semantic encoder with a diffusion-based refinement or reconstruction stage. One single-user framework uses a Swin Transformer for joint source–channel coding and a diffusion model for semantic fine-tuning at the receiver, reporting that the generative decoder improves perceptual quality under AWGN and Rayleigh channels (Zhang et al., 2024). A related personalized federated design uses a pre-trained ViT as the generative semantic encoder and semantic decoder, with channel encoder and decoder blocks placed between them, and couples this stack to Personalized Semantic Federated Learning with Personalized Local Distillation and Adaptive Global Pruning (Peng et al., 2024).

Text-centric designs show a parallel pattern. In KGRAG-SC, the transmitter extracts entities from a sentence, constructs a Minimum Connected Subgraph in a shared multi-dimensional knowledge graph, and transmits only compact node IDs; the receiver reconstructs the subgraph and uses an LLM such as LLaMA 3-8B with a GraphRAG-style prompt to generate the final fluent sentence (Fan et al., 5 Sep 2025). Ti-GSC instead couples a Transformer semantic encoder–decoder with a GAN-based signal distortion suppression module, allowing text semantic communication without CSI by learning to suppress distortion in both syntactic and semantic dimensions (Mao et al., 2023).

Video-oriented systems often use keyframe semantics plus generative completion. One stable-diffusion framework first extracts keyframes and their semantic information, then uses a semantic decoder to reconstruct keyframes and a frame-interpolation module to generate the full video; it further introduces SD-GSC for known-channel denoising and PSD-GSC for unknown-channel denoising (Li et al., 28 Feb 2025). In satellite video transmission, a different design integrates a pre-trained video encoder, an LDPC encoder, and a generative video model fine-tuned through in-context low-rank adaptation, explicitly targeting very high error rates (Zhao et al., 28 Apr 2026).

3. Semantic carriers, coding strategies, and channel adaptation

A defining property of GSC is the replacement of raw payloads by semantically sparse carriers. Remote-visual-monitoring systems transmit semantic maps derived from segmentation, sometimes only when a value-of-information metric or semantic change degree exceeds a threshold (Yang et al., 2023). Agent-driven variants add reinforcement learning to decide whether to transmit the current map, using a state that includes time since last transmission, map size, past semantic changes, and average channel gain (Yang et al., 2024).

Other systems adopt different carriers for the same purpose. In metaverse construction, the semantic encoder outputs keypoints through an hourglass network, and the semantic decoder uses conditional Stable Diffusion plus NeRF to reconstruct the scenery; an OT-enabled semantic denoiser corrects hard keypoints under wireless noise (Wang et al., 2024). In 3D semantic communication, the transmitter uses SAM and NeRF to extract a goal-oriented object and represent it as multi-perspective semantic images; a dual-head adaptive semantic compression model then prunes redundant latent dimensions through a learned mask (Jiang et al., 2024).

Knowledge-grounded systems push semantic sparsification further. KGRAG-SC performs NER, community-guided entity linking, one-hop MCSG construction, and NodeID transmission. It also defines graph-theoretic importance scores from degree and betweenness centrality and applies unequal error protection so that high-importance nodes receive stronger FEC, such as a rate-1/2 convolutional code with 16QAM (Fan et al., 5 Sep 2025). In AIGC provisioning, DeKA-g transmits prompt text and latent representations rather than high-dimensional generated content, and treats compression rate and SNR adaptation as explicit knowledge-alignment problems addressed by MAKD and VGSA (Hu et al., 24 Jun 2025).

Channel adaptation is likewise semanticized. Adaptive Global Pruning in personalized federated GSC uses the uplink SNR to determine a pruning ratio for the aggregated student model, reducing communication load and energy (Peng et al., 2024). FAST-GSC treats latency as a scheduling problem: it parallelizes semantic extraction at the transmitter with diffusion inference at the receiver, then uses reinforcement learning to order semantic-unit extraction and sequential conditional denoising to incorporate late-arriving semantic units (Wang et al., 2024). A plausible implication is that, in GSC, “channel coding” increasingly includes semantic prioritization, model adaptation, and inference scheduling rather than only symbol-level redundancy allocation.

4. Knowledge bases, graphs, and agentic control

Knowledge structure is one of the most distinctive organizing principles in GSC. A knowledge-base-enabled formulation partitions the semantic knowledge base into source, task, and channel KBs. The source KB stores semantic metalets or embedding vectors, the task KB converts commands into task tokens through a task knowledge graph, and the channel KB predicts channel knowledge from location and environmental features (Ren et al., 2023). In this view, the transmitted codeword may include selected metalet indices plus a residual, while the receiver reconstructs or generates content from the recovered indices and residual (Ren et al., 2023).

A related multi-task image system uses a Task KB and a Source KB at both transmitter and receiver. The Task KB maps a natural-language task request to a stored instruction using Sentence-BERT cosine similarity, and the Source KB selects hierarchical Swin-Transformer features for reconstruction or segmentation accordingly; the reconstruction branch then uses a diffusion JSCC decoder (Yuan et al., 2024). This architecture makes task conditioning explicit and keeps the transmitted representation aligned with the downstream objective.

Knowledge graphs provide a more explicit semantic skeleton. KGRAG-SC defines a shared graph

$G=(V,E,\mathcal D,\mathcal C),$

where nodes, directed relation triples, textual descriptions, and communities are all available at both ends. The transmitter constructs a compact semantic subgraph and the receiver uses GraphRAG-style prompting to recover text from partial graph evidence (Fan et al., 5 Sep 2025). The paper explicitly presents this as an interpretable alternative to end-to-end deep learning frameworks that “lack interpretability and struggle with robust semantic selection and reconstruction under noisy conditions” (Fan et al., 5 Sep 2025).

Agentic control appears in several variants. A-GSC uses a reinforcement-learning sampling agent with cross-modality capability and a decoder that combines semantic map prediction with diffusion-based scene generation (Yang et al., 2024). FAST-GSC uses PPO to sequence semantic extraction relative to a denoising deadline and reports a receiver-side semantic-difference module for late-arriving guidance (Wang et al., 2024). The LLM-native framework generalizes this further by casting both sides as AI agents, with “AI brains” that understand and regenerate content (Ren et al., 2024).

5. Applications and empirical performance

The application range reported so far includes text, images, video, metaverse construction, 3D scene transmission, federated edge intelligence, satellite networking, and AIGC provisioning (Ren et al., 2024, Jiang et al., 2024, Guo et al., 11 Aug 2025). Reported gains depend strongly on modality, task metric, channel model, and semantic carrier.

Scenario	Reported result	Citation
Point-to-point video retrieval	$93.03\%$ retrieval accuracy, 36 Kbits overhead, and $\Delta\approx99.98\%$ overhead reduction versus 219 Mbits for traditional communication	(Ren et al., 2024)
Knowledge-graph text communication	semantic similarity $\sim 0.78$ at 4 dB versus 0.29 for a Huffman + 16QAM baseline; $\approx 600$ bits/sentence versus $\approx 1600$ for Huffman and $\approx 2800$ for raw ASCII	(Fan et al., 5 Sep 2025)
Personalized federated GSC on CIFAR-10	global classification accuracy $78.2\%$ versus FedAvg $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$0; communication load reduced by $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$1; median uplink energy $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$2 J versus $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$3 J	(Peng et al., 2024)
FAST-GSC	CLIP score 0.67 versus 0.68 for conventional GSC, with residual task latency reduced from 24 to 11 denoising steps	(Wang et al., 2024)
Single-user image GSC	PSNR improved by $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$4 in AWGN and by $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$5 in Rayleigh compared to CNN-based DeepJSCC	(Zhang et al., 2024)
Wireless video transmission via SD-GSC	PSNR improved by $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$6, $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$7, $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$8, and $$d_{\mathrm{sem}}(\boldsymbol{z},\hat{\boldsymbol{z}}) = \frac{\|\boldsymbol{z}-\hat{\boldsymbol{z}}\|^2}{\|\boldsymbol{z}\|^2},$$9 over ADJSCC, Latent-Diff DNSC, DeepWiVe, and DVST; MSE reduced by $G=(V,E,\mathcal D,\mathcal C),$0, $G=(V,E,\mathcal D,\mathcal C),$1, $G=(V,E,\mathcal D,\mathcal C),$2, and $G=(V,E,\mathcal D,\mathcal C),$3	(Li et al., 28 Feb 2025)
Metaverse construction	latency reduced by $G=(V,E,\mathcal D,\mathcal C),$4, with object status accuracy and viewing experience improved by $G=(V,E,\mathcal D,\mathcal C),$5 and $G=(V,E,\mathcal D,\mathcal C),$6	(Wang et al., 2024)

Additional studies broaden the empirical picture. DeKA-g reports a $G=(V,E,\mathcal D,\mathcal C),$7 improvement in alignment between edge-generated and cloud-generated images, $G=(V,E,\mathcal D,\mathcal C),$8 higher efficiency for compression-rate adaptation, and a $G=(V,E,\mathcal D,\mathcal C),$9 low-SNR performance gain (Hu et al., 24 Jun 2025). In mega-satellite networking, a GSC-aware architecture reduces total bandwidth by $93.03\%$0 across four routing types, at the cost of $93.03\%$1–$93.03\%$2 higher delay in cases requiring extra hops to AI nodes (Guo et al., 11 Aug 2025). In satellite-relay video transmission, a generative design achieves 2.5 dB higher video PSNR than conventional semantic communication at an error rate of $93.03\%$3, and remains robust when the error rate exceeds $93.03\%$4 (Zhao et al., 28 Apr 2026).

These results do not define a single universal benchmark. Instead, they show that GSC gains are task-specific and metric-specific: retrieval accuracy, BLEU, CLIP score, IoU, PSNR, FVD, LPIPS, PIQE, DINO+CLIP alignment, energy consumption, and latency all appear as primary objectives in different systems (Mao et al., 2023, Yuan et al., 21 Apr 2025).

6. Limitations, terminology, and open directions

Several recurring limitations are explicit in the literature. Diffusion-based decoders often incur substantial latency: early remote-monitoring work notes that DDPMs require hundreds to thousands of denoising steps, and video GSC based on Stable Diffusion reports complexity dominated by the SD steps with $93.03\%$5 (Yang et al., 2023, Li et al., 28 Feb 2025). Edge deployment is also a major constraint: one survey lists LLM parameter counts of $93.03\%$6–$93.03\%$7 and recommends pruning, distillation, mixture-of-experts, and compact scenario-customized LLMs (Ren et al., 2024).

Knowledge alignment is another structural challenge. DeKA-g formulates both a generation-knowledge gap between cloud-GAI and edge-GAI and a transmission-knowledge gap between the codec’s assumed channel and the actual channel (Hu et al., 24 Jun 2025). Related works raise KB consistency, dynamic KB update protocols, and cooperative KB evolution as prerequisites for practical deployment (Ren et al., 2023, Fan et al., 5 Sep 2025). This suggests that GSC systems with shared priors are only as robust as their synchronization mechanisms.

Privacy, security, and hallucination control remain open. Reported concerns include leakage via memorized training data, poisoning of semantic embeddings, eavesdropping of prompts, privileged-information leakage in semantic payloads, and the need for semantic integrity checks (Ren et al., 2024, Fan et al., 5 Sep 2025, Guo et al., 11 Aug 2025). AGI-driven GSC also highlights trustworthiness, uncertainty quantification, and fairness as unresolved issues, especially when the receiver must produce human-facing content rather than hidden task logits (Yuan et al., 21 Apr 2025).

A separate source of confusion is terminological. In much of the literature, GSC denotes Generative Semantic Communication; however, some papers use the same acronym for Goal-oriented Semantic Communication, including works on metaverse construction and video transmission via generative AI (Wang et al., 2024, Li et al., 28 Feb 2025). The overlap is not merely lexical: many “goal-oriented” systems are generative at the decoder. A plausible implication is that the field is still consolidating its taxonomy.

Recent work on Schrödinger Bridge-based GSC directly targets two of the most persistent criticisms of diffusion-heavy GSC—hallucination and sampling cost—by replacing Gaussian-to-image transport with an optimal semantic-to-image bridge. The reported gains are at least $93.03\%$8 in FID, $93.03\%$9 in SSIM, and over 8 times faster inference (Gao et al., 20 Apr 2026). Whether such formulations become a general backbone for future GSC remains open, but the direction is consistent with the broader trajectory of the field: tighter coupling between semantic structure, generative priors, and communication constraints.